Verified does not mean true.
Good research states what it measured, how it adjusted, how much the estimate could move, and what it cannot answer. This page sets out that discipline — and draws a hard line between what PolitiMind can prove about a number and what remains statistical inference.
Four standards, not one
Accuracy, provenance, uncertainty, and integrity are separate questions
Most arguments about polling collapse four different standards into a single word — "credible". They are not the same, they are not assured the same way, and only some of them can be proven. Provenance and integrity are claims about the production chain, and an outsider can check them. Accuracy and statistical uncertainty are matters of inference, and they carry error that no method, panel, or ledger removes.
The matrix below is the governing frame for everything else on this page. Read it as a boundary: it is deliberately explicit about what verification does not buy you.
What each standard settles, and what it never settles
Accuracy
Live· available nowDoes this finding match what the population actually thinks?
Estimated, with uncertainty
Statistical inference. It carries error that no method removes.
What it settlesNothing on its own. Accuracy is the target of the whole exercise, approached through sampling design, weighting, and inference — and it is always an estimate.
What it never settlesIt is never established by verification, anchoring, or disclosure. A perfectly verified, perfectly recorded study can still be wrong about the population.
Provenance
Live· available nowWhere did this number come from, and can that origin be reconstructed?
Proven from the record
An outsider can check this without trusting us.
What it settlesThat the figure traces to a specific frozen questionnaire, a specific response set, and a specific recorded computation run against a versioned method.
What it never settlesWhether the design or the method was the right choice. Provenance tells you exactly what was done; it does not grade the decision.
Statistical uncertainty
Live· available nowHow much could this estimate move for sampling reasons alone?
Estimated, with uncertainty
Statistical inference. It carries error that no method removes.
What it settlesThe range the estimate should be read as, given the sample. It makes the imprecision explicit instead of hiding it behind a single number.
What it never settlesIt does not cover every source of error, and a stated interval is not a promise that the true value sits inside it.
Integrity
Beta· tamper-evidence on testnetWas this produced the way the publisher says, and has the record changed since?
Proven from the record
An outsider can check this without trusting us.
What it settlesThat participation was verified, the specification was frozen before fielding, the computation was recorded, and later alteration of the published record is detectable.
What it never settlesWhether the finding is accurate. Integrity is a claim about the production chain and nothing more; it is the strongest thing we can prove and the least it would be honest to imply.
The boundaryVerification, disclosure, and tamper-evidence make a finding checkable and defensible. They do not make it true. A study can be fully verified, fully recorded, and still wrong about the population — which is exactly why the uncertainty stays on the page instead of being argued away.
Readiness labels apply to the assurance surface, not to the standard itself. Tamper-evidence is beta and currently testnet-stage; PolitiMind is not independently audited or third-party certified.
Statistical uncertainty
An estimate is a range, and subgroups are wider than the headline
The most common misreading of a poll is treating a point estimate as a fact and applying the full-sample margin of error to every cut inside it. Smaller subgroups carry wider intervals, and a difference that looks decisive at the top line can be indistinguishable from noise one cut down. We publish the interval at the level being read.
Values in percentage points. Figures are illustrative of interval width, not published PolitiMind results.
- Full sample (n ≈ 1,000)48 percentage points ± 3.1
- Subgroup at one quarter of the sample51 percentage points ± 6.2
- Small subgroup at one tenth of the sample44 percentage points ± 9.8
| Reading | Estimate (percentage points) | Plausible range (percentage points) |
|---|---|---|
| Full sample (n ≈ 1,000) | 48 | 44.9 – 51.1 |
| Subgroup at one quarter of the sample | 51 | 44.8 – 57.2 |
| Small subgroup at one tenth of the sample | 44 | 34.2 – 53.8 |
How to read itRead each row as the whole band, not the centre line. Two figures whose bands overlap heavily should not be reported as a lead, and the same sample supports far weaker statements about its subgroups than about its total.
What the interval does not coverSampling error only. It does not cover nonresponse bias, coverage gaps in the frame, question wording effects, mode effects, or respondents who misreport — which is why limitations are stated separately rather than folded into the interval.
The chain
Every stage between a decision and a published number
Methodology is not a paragraph appended after fielding. It is the sequence that produced the figure, and each stage is either disclosed or it is not. These are the stages a PolitiMind study moves through, in order.
Stages of research production, from decision to published result
- 01
Define the decision
The research question is written as a decision to be informed.
The population it concerns and the precision that would actually change the decision are fixed before anything is fielded.
- 02
Design and freeze the instrument
Wording, order, and response options are reviewed, then versioned and frozen.
The review targets leading language and forced choices. Freezing happens before the first response, so the published wording is the fielded wording.
- 03
Declare the sampling frame
The eligible population and intended quotas are stated up front.
The sample is then judged against a declared frame rather than against whoever happened to answer.
- 04
Field to verified participants
Responses come from participants verified as real, unique, and eligible.
Identity is held apart from the answers themselves, so verification never becomes a record of who said what.
- 05
Apply integrity controls
Duplicate, coordinated, and low-quality participation is flagged before figures are computed.
Controls reduce named failure modes. They do not certify a clean file, and the specific signals are not published.
- 06
Weight to the declared frame
Responses are weighted using a registered, versioned weighting method.
The same inputs reproduce the same output. Study-specific targets are published with the study.
- 07
Compute with a versioned method
Figures come from a recorded computation run citing the method version.
That replaces an untracked spreadsheet nobody outside the vendor can re-execute.
- 08
Publish with uncertainty and limits
The result travels with its frame, weighting approach, interval, and limitations.
Anything the method cannot answer is named by the study itself rather than discovered by a critic.
Disclosure discipline
What gets published, versus what the industry treats as proprietary
The difference is rarely statistical sophistication. It is what a reader is allowed to see. Where legacy practice keeps the mechanics private and asks for trust, we publish enough of the chain that a hostile reviewer can locate the weak point themselves.
Disclosure discipline compared with conventional research practice
| Dimension | Conventional research | PolitiMind |
|---|---|---|
| Sampling frame | Described loosely, if at all, and often reconstructed after publication when the number is challenged. | Declared before fielding and published with the result, so the sample is judged against the frame the study committed to. |
| Questionnaire | Sometimes released, sometimes summarised, with no way to tell whether wording changed mid-field. | Versioned and frozen before the first response; the published wording is the fielded wording. |
| Weighting | Treated as proprietary craft. Outsiders cannot tell which adjustment produced the headline. | The approach and its version are stated and cited by the result, so the same inputs reproduce the same figures. |
| Uncertainty | A single margin of error in a footnote, frequently applied to subgroups it does not cover. | Intervals reported at the level being read, with subgroup imprecision stated rather than inherited from the full sample. |
| Limitations | Omitted, because disclosure weakens the headline. | Named by the study itself and carried with it, including what the method cannot answer. |
| Recomputation | Impossible from outside. The only defence of a disputed number is the vendor's reputation. | The record behind the figure is inspectable, and a published commitment to it can be recomputed by an outsider. |
Comparison describes common industry practice in general terms and names no specific vendor. Weighting targets and integrity thresholds remain study-specific.
Known limits
What this methodology cannot fix
Verified participation removes fabricated and duplicate respondents. It does not manufacture coverage of people who never join a research population, and it does not settle whether a verified sample mirrors the country. Weighting narrows that gap against a declared frame; it does not close it, and aggressive weighting on thin cells makes an estimate less stable, not more authoritative.
Some questions are also badly suited to survey measurement at all — low-salience issues, hypothetical future behaviour, and anything respondents have an incentive to misreport. When that is the case, the honest output is a stated limitation, not a confident number.
What we claim about methodology, and what we do not
- Methodology is stated in plain language — sampling frame, weighting approach, margin of error, and known limitations travel with every result.
- Published results carry provenance: the specification, method version, and computation run behind a number are recorded and inspectable.
- Every study runs behind integrity controls that detect duplicate, coordinated, and low-quality participation before results are published.
- BetaThe platform supports advanced research designs including trackers, recontact studies, experiments, MaxDiff, conjoint, and pre/post measurement.
- BetaResult commitments are anchored to a public ledger so the record is tamper-evident after publication.
PolitiMind claims disclosure and provenance: a declared sampling frame, a frozen instrument, a stated and versioned weighting approach, reported uncertainty, and named limitations travelling with every result. It does not claim to eliminate sampling, nonresponse, coverage, or measurement error, that any sample is representative at every cut, or that verification and anchoring establish accuracy. Tamper-evidence is beta and testnet-stage, and no part of this methodology is independently audited or third-party certified.
Check the discipline
Methodology is stated in plain language — sampling frame, weighting approach, margin of error, and known limitations travel with every result.
Read a published resultPublished results carry provenance: the specification, method version, and computation run behind a number are recorded and inspectable.
See what a record provesThe platform supports advanced research designs including trackers, recontact studies, experiments, MaxDiff, conjoint, and pre/post measurement.
Research methods availableWhy it matters to you
Stated limits are what makes a number survive contact with an opponent
A finding published without its frame, its interval, and its limitations is easy to attack and impossible to defend, because the first credible objection is also the first time anyone has heard the caveat. A finding published with them has already conceded the ground worth conceding, and what remains is the evidence.
That is the trade this methodology makes: less certainty in the headline, far more of it in the decision.
Continue the argument
Proof of Integrity
Inspect how a published number can be checked
Read itWhy Polling Is Broken
Name the structural failure they have felt but not articulated
Read itPolling Platform
See how a study goes from specification to fielded, closed, and reported
Read itPublic Results
Inspect a real published result and its methodology
Read it